Gaussian Processes for Modeling with Uncertainty

Gaussian Processes, like Bayesian methods introduce natural uncertainty to modeling.   And provide more transparency than deep learning neural nets to the models being built and used.  Already being used by many practitioners.   Article provides a good, largely non-technical introduction  and examples of their current use.

In Wired: AI is About to Learn more like Humans, with a little Uncertainty   by Cade Metz

A broader view of artificial intelligence (AI) seeks to equip deep-learning neural networks to better deal with uncertainty via a Bayesian approach that feeds new evidence into existing models, performing functions at which neural networks do not excel. "There are problems in the domain of language and in driverless cars where you're never going to have enough data to use brute force the way that deep learning does," says Geometric Intelligence founder Gary Marcus. Startups are designing neural networks along a Gaussian process (GP) of statistical modeling to identify uncertainty. "Knowing that you don't know is a very good thing," says University of Edinburgh researcher Chris Williams. "Making a confident error is the worst thing you can do." One company is applying GPs toward building AI systems that can learn to navigate massively multiplayer games and other digital environments, with a long-term goal of eventual real-world navigation capability. ... " 

Gaussian Process in WP.

Jobs Susceptible to Computerization

Working on a related problem.  What jobs or part of jobs are likely to be replaced vs augmented?  What is the influence of Cognitive methods?

The Future of Employment: How Susceptible are Jobs to Computerization.
Carl Benedikt Frey†and Michael A. Osborne‡ September 17, 2013
Abstract
We examine how susceptible jobs are to computerisation. To assess this, we begin by implementing a novel methodology to estimate the probability of computerisation for 702 detailed occupations, using a Gaussian process classifier. Based on these estimates, we examine expected impacts of future computerisation on US labour market outcomes, with the primary objective of analysing the number of jobs at risk and the relationship between an occupation’s probability of computerisation, wages and educational attainment. .... " 

The Black Swan

Have read Nassim Nicholas Taleb's book: The Black Swan: The Impact of the Highly Improbable. Highly recommended for modelers or those who think about their application.

Taleb, once a very successful derivative and options trader, now a professor at the University of Massachusetts, takes you on a wild and often idiosyncratic ride through financial modeling. No equations in the book, but it helps to have some basic statistics and econometric background. As close to a page-turner as a book like this can be.

Taleb writes about what he considers the total inadequacy of currently used modeling methods. This is mostly a full-steam attack on the mis-use of Gaussian methods (the Normal or Bell curve), which are the basis of modern portfolio theory, forecasting, regression and just about any statistical method that claims to be predictive. The Gaussians' small tails make it incapable of modeling anything even close to improbable. Our connected world is getting more improbable, thus these methods are increasingly wrong.

To be clear, it's not that the use of Gaussian methods are always wrong, though Taleb's style sometimes implies that. He is making the case that they have been used to underpin all of financial modeling methods, even when it makes no sense.

Along the way, Taleb tells a personal story (very unusual for a statistics book!) and trashes modern portfolio theory, Black-Scholes, Wharton, the Nobel Prize Committee, the use of narrative and most of the last twenty years of econometrics. He has received threats from the normally staid econometric community. He suggests that that Mandelbrot's scalable fractal methods are a better approach than Gaussian methods.

He says that modern practitioners of financial methods agree with him, though most academics do not. The formal methods do not work. Evidenced by the 1998 LTCM crisis, which came close to bringing down the entire global financial system. The formal methods could not deal with a 'black swan', a very improbable event by normal distribution standards.

The practitioners respond that the current portfolio risk methods are all they have today to create useful models. Mandelbrot's models do not give the predictions that current formal methods do. But if the predictions are wrong?

This is a big deal to large companie who use many techniques like forecasting, marketing mix and simulation models that are based on what Taleb is saying are flawed fundamentals in a world of increasing improbabilities. Also, WSJ Review.